YOLO-PEST: a novel rice pest detection approach based on YOLOv5s

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate rice pest detection is vital for intelligent agricultural systems.
• Challenges include limited datasets, pest occlusion, and poor small object detection accuracy.

Purpose of the Study:

• To develop an innovative detection approach, YOLO-PEST, to overcome current limitations in rice pest monitoring.
• To enhance the accuracy and robustness of pest detection models.

Main Methods:

• YOLO-PEST utilizes the YOLOv5s architecture with integrated ConvNeXt module for multiscale feature extraction.
• CoTAttention mechanism is incorporated to improve robustness in complex environments.
• Data augmentation includes random cropping to simulate pest occlusion.

Main Results:

• YOLO-PEST achieved a mean Average Precision (mAP@0.5) of 97%.
• This represents a 1.4-point improvement over previous methods.
• The model demonstrated enhanced accuracy in detecting small and occluded rice pests.

Conclusions:

• YOLO-PEST effectively addresses challenges in rice pest detection, including occlusion and small object identification.
• The proposed method significantly improves detection performance, offering a valuable tool for intelligent agricultural systems.
• The integration of ConvNeXt and CoTAttention enhances model accuracy and robustness.